Skip to content

可以直接跳过微调启动模型,查看效果吗?希望可以更详细的步骤介绍 #112

Description

@bigsmartbig
Image

Activity

  1. bigsmartbig commented on Sep 28, 2025

    @bigsmartbig
    Author

    data_stats_path是指什么的路径?传入前面下载的LIBERO数据集路径可以吗

  2. ModiShi commented on Sep 28, 2025

    @ModiShi
    Contributor
    1. If you have AgiBot G1, you can directly deploy GO-1. Otherwise, you need to finetune on your dataset first.
    2. When finetuning GO-1, data_stats is automatically saved alongside your checkpoint. Pass this file when deploying.
  3. bigsmartbig commented on Sep 29, 2025

    @bigsmartbig
    Author

    Could you explain what each parameter in the payload? Or where can I learn about them?

  4. ModiShi commented on Sep 29, 2025

    @ModiShi
    Contributor

    By default, the model takes three images, one instruction, and robot state (proprioception) as input to output actions. Different LeRobot datasets use different image key names (e.g., "image", "head_color"), which need to be specified in the config. You can also use fewer than 3 images, as shown in our Libero example. The ctrl_freqs parameter sets the robot control frequency - use the default value of 30 if unsure.

  5. simplew2011 commented on Oct 11, 2025

    @simplew2011
    payload = {
        "top": np.array(Image.open("frames/head_color.mp4_20251011_194851.031.png").convert("RGB")),
        "right": np.array(Image.open("frames/hand_right_color.mp4_20251011_194816.624.png").convert("RGB")),
        "left": np.array(Image.open("frames/hand_left_color.mp4_20251011_194739.489.png").convert("RGB")),
        "instruction": "Hanging clothes with a hanger",
        "state": np.expand_dims(np.random.rand(16), axis=0),
        "ctrl_freqs": np.array([30]),
    }
    
  6. zbzyjya commented on Dec 8, 2025

    @zbzyjya
    payload = {
        "top": np.array(Image.open("frames/head_color.mp4_20251011_194851.031.png").convert("RGB")),
        "right": np.array(Image.open("frames/hand_right_color.mp4_20251011_194816.624.png").convert("RGB")),
        "left": np.array(Image.open("frames/hand_left_color.mp4_20251011_194739.489.png").convert("RGB")),
        "instruction": "Hanging clothes with a hanger",
        "state": np.expand_dims(np.random.rand(16), axis=0),
        "ctrl_freqs": np.array([30]),
    }
    

    Thank you for your response. If I want to directly test the effect of the original model, how should I implement video stream input using the payload example you provided, as well as the sequential execution of the robot after the action is returned? Are there any open-source projects or modules that I can refer to?

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions